Scalecast

About

Scalecast helps you forecast time series. It is especially good in the following use cases:

• Series need dynamic transformations to extract signals
• Series have missing values
• Need to establish proof-of-concept experimentations with model validation
• Academic research
• Working in internal production and sandbox environments

Here is how to initiate its main object:

from scalecast.Forecaster import Forecaster

f = Forecaster(
    y = array_of_values,
    current_dates = array_of_dates,
    future_dates=fcst_horizon_length,
    test_length = 0, # do you want to test all models? if so, on how many or what percent of observations?
    cis = False, # evaluate conformal confidence intervals for all models?
    metrics = ['rmse','mape','mae','r2'], # what metrics to evaluate over the validation/test sets?
)

Uniform ML modeling (with models from a diverse set of libraries, including scikit-learn, statsmodels, and tensorflow), reporting, and data visualizations are offered through the Forecaster and MVForecaster interfaces. Data storage and processing then becomes easy as all applicable data, predictions, and many derived metrics are contained in a few objects with much customization available through different modules. Feature requests and issue reporting are welcome! Don't forget to leave a star!⭐

Documentation

• Read the Docs
• Introductory Notebook
• Change Log

Popular Features

1. Easy LSTM Modeling: setting up an LSTM model for time series using tensorflow is hard. Using scalecast, it's easy. Many tutorials and Kaggle notebooks that are designed for those getting to know the model use scalecast (see the aritcle).

f.set_estimator('lstm')
f.manual_forecast(
    lags=36,
    batch_size=32,
    epochs=15,
    validation_split=.2,
    activation='tanh',
    optimizer='Adam',
    learning_rate=0.001,
    lstm_layer_sizes=(100,)*3,
    dropout=(0,)*3,
)

2. Auto lag, trend, and seasonality selection:

f.auto_Xvar_select( # iterate through different combinations of covariates
    estimator = 'lasso', # what estimator?
    alpha = .2, # estimator hyperparams?
    monitor = 'ValidationMetricValue', # what metric to monitor to make decisions?
    cross_validate = True, # cross validate
    cvkwargs = {'k':3}, # 3 folds
)

3. Hyperparameter tuning using grid search and time series cross validation:

from scalecast import GridGenerator

GridGenerator.get_example_grids()
models = ['ridge','lasso','xgboost','lightgbm','knn']
f.tune_test_forecast(
    models,
    limit_grid_size = .2,
    feature_importance = True, # save pfi feature importance for each model?
    cross_validate = True, # cross validate? if False, using a seperate validation set that the user can specify
    rolling = True, # rolling time series cross validation?
    k = 3, # how many folds?
)

4. Plotting results: plot test predictions, forecasts, fitted values, and more.

import matplotlib.pyplot as plt

fig, ax = plt.subplots(2,1, figsize = (12,6))
f.plot_test_set(models=models,order_by='TestSetRMSE',ax=ax[0])
f.plot(models=models,order_by='TestSetRMSE',ax=ax[1])
plt.show()

5. Pipelines that include transformations, reverting, and backtesting:

from scalecast import GridGenerator
from scalecast.Pipeline import Transformer, Reverter, Pipeline
from scalecast.util import find_optimal_transformation, backtest_metrics

def forecaster(f):
    models = ['ridge','lasso','xgboost','lightgbm','knn']
    f.tune_test_forecast(
        models,
        limit_grid_size = .2, # randomized grid search on 20% of original grid sizes
        feature_importance = True, # save pfi feature importance for each model?
        cross_validate = True, # cross validate? if False, using a seperate validation set that the user can specify
        rolling = True, # rolling time series cross validation?
        k = 3, # how many folds?
    )

transformer, reverter = find_optimal_transformation(f) # just one of several ways to select transformations for your series

pipeline = Pipeline(
    steps = [
        ('Transform',transformer),
        ('Forecast',forecaster),
        ('Revert',reverter),
    ]
)

f = pipeline.fit_predict(f)
backtest_results = pipeline.backtest(f)
metrics = backtest_metrics(backtest_results)

6. Model stacking: There are two ways to stack models with scalecast, with the StackingRegressor from scikit-learn or using its own stacking procedure.

from scalecast.auxmodels import auto_arima

f.set_estimator('lstm')
f.manual_forecast(
    lags=36,
    batch_size=32,
    epochs=15,
    validation_split=.2,
    activation='tanh',
    optimizer='Adam',
    learning_rate=0.001,
    lstm_layer_sizes=(100,)*3,
    dropout=(0,)*3,
)

f.set_estimator('prophet')
f.manual_forecast()

auto_arima(f)

# stack previously evaluated models
f.add_signals(['lstm','prophet','arima'])
f.set_estimator('catboost')
f.manual_forecast()

7. Multivariate modeling and multivariate pipelines:

from scalecast.MVForecaster import MVForecaster
from scalecast.Pipeline import MVPipeline
from scalecast.util import find_optimal_transformation, backtest_metrics
from scalecast import GridGenerator

GridGenerator.get_mv_grids()

def mvforecaster(mvf):
    models = ['ridge','lasso','xgboost','lightgbm','knn']
    mvf.tune_test_forecast(
        models,
        limit_grid_size = .2, # randomized grid search on 20% of original grid sizes
        cross_validate = True, # cross validate? if False, using a seperate validation set that the user can specify
        rolling = True, # rolling time series cross validation?
        k = 3, # how many folds?
    )

mvf = MVForecaster(f1,f2,f3) # can take N Forecaster objects

transformer1, reverter1 = find_optimal_transformation(f1)
transformer2, reverter2 = find_optimal_transformation(f2)
transformer3, reverter3 = find_optimal_transformation(f3)

pipeline = MVPipeline(
    steps = [
        ('Transform',[transformer1,transformer2,transformer3]),
        ('Forecast',mvforecaster),
        ('Revert',[reverter1,reverter2,reverter3])
    ]
)

f1, f2, f3 = pipeline.fit_predict(f1, f2, f3)
backtest_results = pipeline.backtest(f1, f2, f3)
metrics = backtest_metrics(backtest_results)

8. Transfer Learning: Train a model in one Forecaster object and use that model to make predictions on the data in a separate Forecaster object.

f = Forecaster(...)
f.auto_Xvar_select()
f.set_estimator('xgboost')
f.cross_validate()
f.auto_forecast()

f_new = Forecaster(...) # different series than f
f_new = infer_apply_Xvar_selection(infer_from=f,apply_to=f_new)
f_new.transfer_predict(transfer_from=f,model='xgboost') # transfers the xgboost model from f to f_new

Installation

Required Installations

• UV recommended
• Only the base package is needed to get started:
  • pip install --upgrade scalecast
  • uv pip install --upgrade scalecast (recommended)

Optional Installations

• shap: feature importance (known issue with Python 3.11+)
• tf: tensorflow for rnn/lstm models (for MAC, you may need to run uv pip install tensorflow-macos tensorflow-metal)
• darts: theta
• greykite: silverkite model
• prophet: prophet model
• tbats: tbats
• lgbm: lightgbm

Install these by using

uv pip install "scalecast[list_optional_dependencies]"

For example, install tensorflow and darts using:

uv pip install "scalecast[tf,darts]"

Please note that the optional dependencies may not be tested before new releases.

Papers that use scalecast

• Post-covid customer service behavior forecasting using machine learning techniques
• Application of ANN and traditional ML algorithms in modelling compost production under different climatic conditions
• Reservoir Computing Solutions for Streamflow Modeling and Prediction in Real World Scenarios
• LSTM-based recurrent neural network provides effective short term flu forecasting
• IMPLEMENTING AN ENERGY TRADING STRATEGY USING FORECASTING OF ENERGY PRICES AND PRODUCTION
• Modelamiento predictivo del número de visitantes en un centro comercial

Udemy Course

Scalecast: Machine Learning & Deep Learning

Blog posts and notebooks

Forecasting with Different Model Types

• Sklearn Univariate
  • Expand your Time Series Arsenal with These Models
  • Notebook
• Sklearn Multivariate
  • Multiple Series? Forecast Them together with any Sklearn Model
  • Notebook 1
  • Notebook 2
• RNN
  • Exploring the LSTM Neural Network Model for Time Series
  • LSTM Notebook
  • RNN Notebook
• ARIMA
  • Forecast with ARIMA in Python More Easily with Scalecast
  • Notebook
• Theta
  • Easily Employ A Theta Model For Time Series
  • Notebook
• VECM
  • Employ a VECM to predict FANG Stocks with an ML Framework
  • Notebook
• Stacking
  • Stacking Time Series Models to Improve Accuracy
  • Notebook
• Other Notebooks
  • Prophet
  • Combo
  • Holt-Winters Exponential Smoothing

Transforming and Reverting

• Time Series Transformations (and Reverting) Made Easy
• Notebook

Confidence Intervals

• Easy Distribution-Free Conformal Intervals for Time Series
• Dynamic Conformal Intervals for any Time Series Model
• Notebook 1
• Notebook 2

Dynamic Validation

• How Not to be Fooled by Time Series Models
• Model Validation Techniques for Time Series
• Notebook

Model Input Selection

• Variable Reduction Techniques for Time Series
• Auto Model Specification with ML Techniques for Time Series
• Notebook 1
• Notebook 2

Scaled Forecasting on Many Series

• May the Forecasts Be with You
• Introductory Notebook Section

Transfer Learning

• Notebook 1
• Notebook 2

Contributing

• Contributing.md
• Want something that's not listed? Open an issue!

How to cite scalecast

@misc{scalecast,
  title = {{scalecast}},
  author = {Michael Keith},
  year = {2024},
  version = {<your version>},
  url = {https://scalecast.readthedocs.io/en/latest/},
}